Coherent forecast combination: A stacked regression approach
Forecast reconciliation is a well-established post-forecasting process that adjusts base forecasts produced by a single expert or model to ensure coherence within a constrained forecasting framework. However, in many practical applications, multiple forecasts are available for each variable, originating from different models or experts. Coherent forecast combination extends traditional reconciliation by integrating forecast combination and reconciliation into a unified framework to improve accuracy and satisfy the coherence property. This methodology includes both optimal (in the least squares sense) and sequential approaches. In the former case, forecast combination and reconciliation occur simultaneously, while in the latter either step is performed separately: first single-variable forecast combination is performed, and then the resulting forecasts are reconciled, or vice versa. In this talk, we explore the theoretical principals of coherent forecast combination, its advantages over single-task combination and single-expert reconciliation approaches, and its practical implementation with the R package FoCo2. Using a real-world energy dataset, we illustrate its effectiveness and its ability to improve forecasting accuracy while fulfilling constraints.
Slide
Download pdf